release-train: develop -> staging

.github/workflows/kanban-archive.yml

@@ -267,6 +267,23 @@ jobs:
           # under suspicion here -- reading the baseline with it would put both
           # sides of the comparison back inside one view.
           GH_TOKEN: ${{ github.token }}
+          # `gh run download` TAKES NO --repo AND THIS JOB HAS NO CHECKOUT, so it
+          # resolved the repository from the git remote, found no `.git`, and died
+          # before making any API call (backend#2802). That is why the artifacts
+          # LISTING above worked -- it puts the repo in the URL path -- while the
+          # fetch failed, and the run went red having archived successfully.
+          #
+          # Reproduced outside Actions, gh 2.98.0:
+          #   $ cd /tmp/empty && gh run download <id> -n board-baseline
+          #   failed to run git: fatal: not a git repository
+          #   $ GH_REPO=tracebloc/.github gh run download <id> -n board-baseline
+          #   $ cat prev/board.total -> 574
+          #
+          # `gh` does NOT read `GITHUB_REPOSITORY` (tested), so being on a runner
+          # does not supply it. GH_REPO is the variable it does read, and it is
+          # cheaper and narrower than adding a checkout to a job that needs no
+          # source.
+          GH_REPO: ${{ github.repository }}
         run: |
           set -uo pipefail
           : > prev.total
@@ -313,7 +330,27 @@ jobs:
             else
               rid=$(printf '%s' "$arts" \
                 | jq -sr '[.[] | select(.expired == false)] | sort_by(.created_at) | last | .workflow_run.id')
-              if gh run download "$rid" -n board-baseline -D prev >/dev/null 2>&1 && [ -s prev/board.total ]; then
+              # KEEP THE DOWNLOAD'S STDERR. `>/dev/null 2>&1` is the reason this
+              # failure was diagnosable only by reproducing it by hand: the step
+              # reported "could not be downloaded" and threw away the one line
+              # that said why ("failed to run git: fatal: not a git repository").
+              # Stdout still goes to /dev/null -- it is progress noise -- but the
+              # error text is captured and echoed by the refusal below.
+              # STDERR TO A SCRATCH FILE, PROMOTED TO prev.error ONLY ON FAILURE
+              # (Bugbot, High -- on my own fix). Writing straight into prev.error
+              # couples the two: the assert step at the bottom treats ANY non-empty
+              # prev.error as an unreadable baseline (`if [ -s prev.error ]`), so a
+              # SUCCESSFUL download that printed anything at all would mark the
+              # baseline unreadable and keep the job red after a good archive --
+              # the exact failure this PR fixes, arriving from the other side.
+              #
+              # MEASURED, and stated honestly: `gh run download` writes 0 bytes to
+              # stderr on success on a non-TTY (tested, gh 2.98.0), so the finding
+              # as described does not reproduce today. The COUPLING is real
+              # regardless, and one deprecation notice on stderr in a future gh is
+              # all it would take. Decoupled unconditionally because it costs
+              # nothing: only the failure branch can now populate prev.error.
+              if gh run download "$rid" -n board-baseline -D prev >/dev/null 2>prev.stderr && [ -s prev/board.total ]; then
                 cp prev/board.total prev.total
                 echo "Baseline: run $rid saw $(cat prev.total) item(s)."
                 # IS THIS JOB THE ONLY ARCHIVER SINCE? It is not, and assuming
@@ -353,7 +390,14 @@ jobs:
                 # this is a broken read, not an empty history. Refusing is the
                 # whole point of having asked the two questions separately.
                 # selftest:unreadable-path
-                printf 'a live board-baseline artifact exists (run %s) but could not be downloaded\n' "$rid" > prev.error
+                # APPEND, DO NOT CLOBBER. `gh run download`'s stderr is already in
+                # prev.error at this point, and it holds the only line that says
+                # WHY -- overwriting it with this summary is what made the last
+                # five red runs undiagnosable from their own logs. The summary
+                # goes first because it names the run; the tool's own words follow.
+                reason=$(tr '\n' ' ' < prev.stderr 2>/dev/null | cut -c1-300)
+                printf 'a live board-baseline artifact exists (run %s) but could not be downloaded. gh said: %s\n' \
+                  "$rid" "${reason:-<no stderr captured>}" > prev.error
                 echo "::warning::the baseline exists but could not be read; the assert step will refuse."
               fi
             fi